Berkson's paradox is a false correlation produced by the way a sample was selected. Two unrelated traits appear related — often negatively — because the selection required at least one of them to be present.
The relationship exists in the sample and nowhere else.
Suppose talent and attractiveness are entirely unrelated in the population, and a casting agency signs anyone who is sufficiently talented or sufficiently attractive.
Among the signed actors, the two traits will be negatively correlated. Someone who got in on talent is unlikely to also be exceptionally attractive — if they were, they would have cleared the bar twice, which is rarer. The agency's books now show a trade-off that does not exist outside them.
The technical name for what has happened is conditioning on a collider — selecting on a variable that both traits influence.
Berkson's paradox is named for a 1946 observation about hospital data: because admission usually requires at least one serious condition, diseases that are unrelated in the general population can appear inversely related among inpatients. The hospital is not a neutral window onto the population; it is a filter, and the filter creates the pattern.
The same mechanism operates anywhere a sample is defined by a threshold. University cohorts selected on a combination of grades and interviews will show grades and interview performance trading off. Dating pools selected on several desirable traits will suggest those traits are in tension. In none of these cases does the correlation exist outside the selected group.
It is harder to spot than most sampling problems because the sample is not obviously unrepresentative — it may be a complete census of the group in question. The question to ask is not “is this sample biased?” but “what did it take to get into this sample, and could the traits I am comparing both have contributed?”